BLA: Blind LLM Agents

BLA: Blind LLM Agents

Hardens your own software builds against AI reverse-engineering agents, which hit resource limits and decoys while your users see the same runtime behavior.

Made by qybaihe on GitHub. These notes are Mindket's, written from its README; the project's own files and license are in its repository.

About

What it is: BLA is a defensive tool that hardens the builds you ship against AI reverse-engineering agents, which the README calls REA. It plants decoys and pushes artifacts past the analyzer's resource budgets, so the analyzer wastes effort and the program runs the same for your users. It is meant for your own software.

Who it suits: It suits developers who ship software and worry about AI agents taking it apart.

Industry | Who uses it | How they use it | Example
Desktop software | Electron app developers | Harden the app's ASAR archive before release | An ASAR header grown from 607 B to 16.9 MiB
Backend and systems | Go developers | Harden release binaries against automated analysis | Go build info grown from 50 KiB to 1.16 MiB
Blockchain | Teams shipping EVM bytecode | Harden the bytecode before it goes out | Bytecode grown from 16 B to 4.00 MiB

Anyone who wants to analyze software they do not own should skip it, since BLA is for protecting your own.

Where to use it: BLA runs from a terminal and from your coding agent.

• Terminal: run the npx commands in your project folder
• Coding agents: setup registers an MCP server and a skill, so you can ask for a pass in plain words
• Build outputs: it finds dist and build folders, .asar, .wasm and .sqlite files, and models

Use case | Who | What they get back
Harden a release | Developers shipping builds | A hardened build whose runtime behavior stays equivalent
Check the defense | Security-minded teams | A 14-target benchmark with local estimates and runtime checks
Ask in plain words | Coding agent users | A hardening pass started from a short prompt
Run stages one by one | Teams that want control | Audit, protect, verify or defend run on their own

How to set it up: You need Node.js and a project with a build before you start.

• Node.js 24.x, or 22.23.4 and up in the 22 line
• A project with build output, such as dist or build
• A coding agent, if you want to ask for hardening in plain words

1. Install a recommended Node.js version.
2. Run setup once, which registers the MCP tools and the skill with your coding agent.
3. Run harden on your project folder.
4. Run the benchmark if you want to see the showcase.

npx bla-agents setup
npx bla-agents benchmark

How to use it: Run harden from your project folder, or ask your agent after setup.

npx bla-agents harden .

Harden my software against REA

Tool | What it does
harden | Finds your build outputs and runs audit, protect and verify in one pass
audit, protect, verify | The three stages of that pass, which you can also run one at a time
defend | Runs the same pipeline when you already know the path
benchmark | Runs a 14-target showcase with local estimates and runtime checks
setup | Registers the MCP tools and the skill with your coding agent
bla_harden | The MCP tool an agent calls to do the same job

Example: After setup, you ask your agent to harden your software against REA. BLA finds your build outputs, then audits, protects and verifies them. An AI reverse-engineering agent that opens the result runs into resource budgets and decoys, while your users see the same runtime behavior.

For newcomers and experts: REA is an AI agent that reverse-engineers release builds with one command, and BLA is built as the counter to it. If you are new, install Node.js, run setup, and try harden on a small test project first.

BLA works in three layers. Poison adds bait metadata, canary values in valid formats and fake control flow paths. Exhaust pushes an artifact past REA's documented budgets for headers, nesting and interface inventories. Hallucinate leaves the analyzer with incomplete inventories, failed redactions and murky dead ends. The trade-off is that the measured results cover only the REA analyzer path in the author's environment, not sessions with Ghidra, Frida, LLDB, Rizin or JADX, and artifacts grow in size.

You point BLA at the build you ship
BLA runs audit, protect and verify in one pass
The hardened build keeps the same runtime behavior
An AI reverse-engineering agent hits budgets and decoys

Good to know:

• Defensive use: the README says BLA is for protecting your own software and links a disclaimer
• Results: measured against real REA on ordinary hardware, covering only the analyzer path that was run
• Benchmark: 14 targets, with internal estimates
• Format coverage: the table below shows Go build info, Erlang BEAM, EVM bytecode, PCAP and Electron ASAR, and its search for build outputs also covers .wasm, .sqlite and model files
• License: MIT

Title: Verified against real REA
Format | Before | After | REA outcome
Go buildinfo | 50 KiB | 1.16 MiB | Capacity error at the 1 MiB limit
Erlang BEAM | 84 B | 262 KiB | Capacity error at the 64-chunk budget
EVM bytecode | 16 B | 4.00 MiB | Approaches the full 4 MiB input cap
PCAP | 24 B | 1.29 MiB | Budget rejected, with tshark present
Electron ASAR | 607 B | 16.9 MiB | Capacity error at the 16 MiB header limit

On GitHub, 6 November 2026

• Stars: 2,486
• Forks: 327
• Watching: 10
• Open issues and pull requests: 4
• Language: TypeScript
• License: MIT
• Topics: adversarial-ai, claude-code, cursor, electron-security, llm-security, mcp, mcp-server, obfuscation, reverse-engineering, webassembly
• Trending: #19 repository of the day on GitHub trending (6 November 2026)

Links

• Code: https://github.com/qybaihe/bla
• Website: https://bla.tools/
• Releases: https://github.com/qybaihe/bla/releases
• Issues: https://github.com/qybaihe/bla/issues
• Star history: https://www.star-history.com/#qybaihe/bla&Date

Get the code

git clone https://github.com/qybaihe/bla.git

Then follow the setup in its README. Nothing in this kit runs by itself.